Engineer / Consultant / Human

I build AI systems
that have to work.

Product engineering, operational systems, and technical problem-solving grounded in clear requirements and verifiable results.

Jeff GraberNew YorkWorking since 1998
See Tenant401 →
Practice
Since 1998
25+ years building and operating systems
Scale
30,000+
Clients. Real people, real stakes.
Enterprise
Viacom · UN · Hilfiger
High-stakes environments
Approach
Spec-first
Define done. Build to it. Verify.
Current system
01 / Tenant dispute platform
Spec-first AI system that helps tenants navigate housing disputes. Deterministic state graph, structured delivery, zero hallucination tolerance.
Year2026
TagAI systems
Tagproduct
StatusWorking product
Artifact
Pending
State graph · Verification output · Interface screenshot
Field Notes
04
AI stays a pilot when nobody defines what working means
Why spec-first engineering matters more than the model you choose.
Note2026-09-22
03
The state graph is the product
Why logic belongs in deterministic code, not in prompts.
Note2026-09-18
02
Thirty thousand edge cases
What running a service business teaches you that engineering school doesn't.
Note2026-09-10
01
What a verification step actually does
On checking AI output before it reaches the user — and why most teams skip it.
Note2026-08-28
How it gets built
Method
01
Define.
Write requirements and criteria for success before touching anything.
02
Constrain.
What can the system never do? What happens when it doesn't know?
03
Build.
State graph first. Language second. A useful thing beats a perfect demo.
04
Verify.
Every output checked against requirements before the user sees it.
05
Operate.
Deployed, monitored, maintained. A system that keeps working.
Brief / Acceptance criteria
SYSTEM: Tenant dispute navigator

MUST:
  — Identify the correct dispute type
  — Cite applicable NYC housing law
  — Surface the correct next action
  — Acknowledge uncertainty explicitly

MUST NEVER:
  — Invent a regulation
  — Provide legal advice
  — Proceed if confidence < threshold

SUCCESS WHEN:
  — User can take a real next step
  — No hallucinated facts in output
Operating History

I learned systems by being responsible for them.

1998
Enterprise infrastructure
Built systems for Viacom, Tommy Hilfiger, and the United Nations through j.u.m.p! inc.
2006
The Mac Support Store
Seventeen years, 30,000+ customers in Brooklyn. Closed deliberately when the era ended.
Now
AI systems with clear contracts
Spec-first delivery. Deterministic logic. Verification before shipping.
Working

Bring me the problem that isn't behaving.

Describe what you're trying to make work →
jeff@jeffgraber.com